Closed
Description
For example the following test code:
extern "C" {
fn random_bool() -> bool;
fn random_usize() -> usize;
fn output(_: usize);
}
#[no_mangle]
pub fn test() {
let condition = unsafe { random_bool() };
let num = unsafe { random_usize() };
for i in 0..num {
if !condition {
break;
}
unsafe { output(i) };
}
}In assembly generated by rustc, it can be seen that the check of condition is inside the loop, while it should be considered a loop invariant, and be pulled out from the loop.
Note that Clang is able to do this optimization with an equivalent C++ code:
#include <cstddef>
bool random_bool();
size_t random_size();
void output(size_t n);
void test() {
auto condition = random_bool();
auto count = random_size();
for (size_t i = 0; i < count; i++) {
if (!condition) {
break;
}
output(i);
}
}so it seems either this is something should be done in compiler frontend, or we are missing some optimization from LLVM.